intersubject differences in volume (fig. S5),
but differences between the two hemispheres
were not significant at a global level, nor
were differences between male and female
brains (see table S2 for full information on these
volumes). The list of volumes is a growing
resource together with the probabilistic maps
of areas and nuclei, made available and con-
stantly updated through the Knowledge graph
of the European Human Brain Project (HBP;
see https://ebrains.eu).
The comparison of the degree of intersub-
ject variability suggests differences between
brain regions (fig. S5). For example, we found
high variability (i.e., low values for probability)
in Broca’s region with areas 44 and 45 and the
superior parietal lobule, whereas the occipital
pole with the primary and secondary visual
cortices (BA17/18) and area Te3 in the tempo-
ral lobe appeared less variable.
Previous and ongoing mapping projects
resulted in more than 10,616 XML files con-taining 85,210 contour lines with 3,737,771 points
and a total length of 1961 m. Stacks with con-
tour lines were managed using the open-source
version control software Subversion, which
automatically manages files and directories
so as to document the complete history of how
the localization of an areal border might have
changed over its life cycle (fig. S1B). Changes
may occur when a new mapping project re-
quires reanalysis of an already existing map,
but these have been small in the past.Amuntset al.,Science 369 , 988–992 (2020) 21 August 2020 3of5
Fig. 2. Workflow of the 3D reconstruction of serial histological sections
and alignment of brain data to a reference space, cytoarchitectonic analysis
in 2D images, and the computing of the probabilistic Julich-Brain atlas.
(A) To recover the 3D shape of a postmortem brain, we applied a combination of
linear and nonlinear processing steps at different scales based on the undistorted
MRI dataset, and optional on blockface images, obtained during sectioning. The
digitized histological images were repaired and corrected for illumination and
optical imbalances. A rigid section-to-section alignment was computed to create
a first approximate 3D reconstruction. It served to align the MRI dataset to
the corresponding section planes by a rigid-body transformation. The sections
were nonlinearly registered to the sections of the MRI by an elastic method.
The alignment was improved by re-registering the output several times, section
by section, to a median filtered version. (B) The cytoarchitecture was analyzed in
consecutive histological sections covering the complete extent of an area and
characterized by the gray-level index ( 16 ). Contour lines of the areas were
submitted to a data repository, 3D-reconstructed, and topologically normalized.
Linear and nonlinear transformations were applied to the areas, and areas
were superimposed to form the cytoarchitectonic probability map. (C) Volume-
and surface-based maximum probability maps of the Julich-Brain atlas were
computed. To effectively organize the intensive computations, we implemented a
data processing management system that allowed distributed processing of a
large number of datasets across multiple CPU cores. It was designed to scale up
well from a single core computer system up to thousands of computing nodes in
a high-performance computing environment ( 12 ).Seefig.S1,AtoC,fordetails.RESEARCH | REPORT